Robustness Analysis Of The Independence Hypothesis Of Panel Data Program Evaluation Method

Hsiao et al.(2012)proposed a new method for program evaluation of panel data in the direction of causal inference,which attributed the correlation between cross-sectional units to the existence of some common factors.Comparing with the previous policy evaluation methods,they simplified the procession of constructing the"Counterfactual",which are unobservable and difficult to estimate,with proving the cross-sectional units can take the role of common factors in their approach.In this paper,we study the robustness of the panel data approach method for program in both theoretical and practical.On the one hand,we provide a sensitivity analysis of independent hypotheses under various data settings,which is one basic theoretical hypothesis about the random error term(random perturbation term)of the cross-sections.And then,a comparison of different variable selection methods in it.It is found that the panel data policy evaluation method is very sensitive to the independence hypothesis in the case of high-dimensional cross-sectional data,which will lead to inaccurate prediction of the "Counterfactual",and influence the estimation of the impact of policy.In this paper,we also propose using the principal component regression instead of the original variable selection process in panel data policy evaluation method to improve the precision.The conclusions of the Monte Carlo simulations also show that even under the situation of relaxing assumption about independence,the out-of-sample prediction error with principal component regression method is smaller than the method of variable selection method in panel data policy evaluation.On the other hand,we provide the evaluation of the Closer Economic Partnership Arrangement(CEPA)for Hong Kong's economy in 2003,with the new method about principal component regression.It is found that the implementation of CEPA has made Hong Kong's actual GDP increase by 2.66%year-on-year,which demonstrated by other different methods,placebo studies and robustness test.